Multi-GPU Compute Scaling: NVLink vs. Infinity Fabric

This article provides a comprehensive analysis of how NVIDIA's NVLink and AMD's Infinity Fabric perform in modern multi-GPU compute benchmarks. It explores why traditional PCIe lanes bottleneck high-performance workloads, how these proprietary high-speed interconnects enable memory pooling, and the actual scaling efficiency achieved in artificial intelligence, rendering, and scientific simulations using today's top-tier hardware.

The Death of PCIe in Heavy Compute

To understand the benchmark performance of NVLink and Infinity Fabric, you must first understand the limitation of standard PCIe slots. Even modern PCIe Gen 5 x16 offers a bidirectional bandwidth of roughly 128 GB/s. In complex compute benchmarks—such as training Large Language Models (LLMs) or running fluid dynamics simulations—GPUs must constantly share massive datasets. When forced through PCIe, GPUs spend more time waiting for data than processing it, resulting in poor multi-GPU scaling. NVLink and Infinity Fabric bypass this bottleneck by creating direct, ultra-high-bandwidth bridges between GPUs.

NVIDIA’s NVLink, particularly in its fourth and fifth generations (found in Hopper H100 and Blackwell B200 architectures), dominates AI and deep learning benchmarks.

  • Bandwidth: Modern NVLink provides up to 900 GB/s (Hopper) and 1.8 TB/s (Blackwell) of bidirectional bandwidth per GPU—over ten times the speed of PCIe Gen 5.
  • Scaling Efficiency: In standard compute benchmarks like MLPerf, moving from one to eight GPUs using standard PCIe often yields dismal scaling efficiency (sometimes dropping below 50% per added card due to communication overhead). With NVLink and NVSwitch, scaling efficiency regularly hits 90% to 95%. This means an 8-GPU cluster performs nearly eight times faster than a single GPU.
  • Memory Pooling: NVLink allows multiple GPUs to pool their VRAM into a single, unified memory space. In rendering benchmarks like OctaneBench or V-Ray, if a scene exceeds a single GPU's memory buffer, performance normally crashes. With NVLink, the workload scales seamlessly across the combined memory of the linked GPUs without a performance penalty.

AMD Infinity Fabric Benchmark Performance

AMD utilizes its Infinity Fabric interconnect technology mainly within its Instinct MI-series accelerators (such as the MI250X and MI300X) to power high-performance computing (HPC) and exascale supercomputers.

  • Bandwidth: The latest Infinity Fabric implementations allow up to 896 GB/s of peer-to-peer bandwidth between GPUs in a standard 8-GPU node.
  • Coherent Memory: AMD's distinct advantage in benchmarking is CPU-to-GPU memory coherency, particularly in the MI300A (which combines CPU and GPU cores) and xGMI links. This eliminates the need to copy data back and forth between system RAM and VRAM, drastically reducing latency in scientific compute benchmarks like LAMMPS (molecular dynamics) and GROMACS.
  • Scaling Efficiency: In FP64 (double-precision) compute benchmarks, AMD’s Infinity Fabric demonstrates near-linear scaling. When scaling from two to eight Instinct GPUs, HPC workloads regularly maintain an 85% to 92% scaling efficiency, trading blows directly with NVIDIA's NVLink in raw compute throughput.

Workload Dependency: Where Scaling Matters

Hardware interconnects only perform well if the benchmarking software is programmed to use them. Multi-GPU scaling falls into two categories during testing:

  1. Weakly Coupled Workloads (PCIe is Enough): Benchmarks like Blender Cycles, Cinebench, or basic cryptocurrency mining split a task into smaller, independent chunks. Each GPU works on its own piece of the puzzle and rarely talks to the others. Here, NVLink and Infinity Fabric show almost zero performance improvement over standard PCIe. Scaling is already near 100% regardless of the interconnect.
  2. Tightly Coupled Workloads (Interconnect is Critical): Deep learning training, weather forecasting, and massive-scale finite element analysis require every GPU to know what the others are calculating millisecond by millisecond. In these benchmarks, NVLink and Infinity Fabric deliver a 2x to 4x total system speedup compared to running the exact same GPUs over a standard PCIe motherboard interface.

The Verdict on Modern Hardware

In modern compute benchmarking, both NVLink and Infinity Fabric have effectively solved the multi-GPU scaling bottleneck for enterprise hardware. While consumer-grade GPUs have largely had these features stripped away (NVIDIA removed NVLink from the RTX 40-series), workstation and data center hardware rely on them entirely. When benchmarking AI and HPC tasks, both technologies deliver 85% to 95% linear scaling, turning multiple discrete graphics cards into what the software perceives as one massive, unified processor.